Tools And Recovery

In consolidated mode, list, summary and prevalidate share list_azure_advisor_recommendations; they are not separately exposed tools. Use the connected schema to select the operation.

For deletion evidence, verdicts and already-deleted resources, follow Read-only prevalidation.

Frequency/count/grouping uses normal summary and requested group-by; both modes return Cost. List and summary use only subscriptions, recommendation-type-ids and resource-types arrays, including one-element arrays for single values. Prevalidation also uses subscriptions but keeps its one required scalar recommendation-type-id. Metadata list keeps scalar resource-type; metadata/remediation get keep scalar recommendation-type-id. Updates retain scalar subscription and their existing single-ID/batch-ID inputs. Do not send removed singular read aliases. Read arrays use OR within each filter; subscription and SG scope cannot combine. minimum-savings is non-negative annual retail USD potential savings, not monthly/contractual. Minimum-savings and summary as-of/from/to require both Cost and IdleResources; dates use UTC ISO 8601; from <= to, no future as-of. List has no historical inputs.

Use the connected schema for supported filter values and limits. Map friendly wording to a supported product family only when unambiguous; no provider mappings. Only send fully qualified resource-type(s) when supplied. For unresolved recognizable names or empty family-filtered results, resolve stable IDs with metadata_list, retry and disclose fallback. Search matches instance problem text, not metadata labels. Preserve product-family with prioritization; filtering precedes totals/ranking. Prioritized summary groups by type; prioritized list requires exactly one distinct type in a recommendation-type-ids array. Summary limits affect displayed groups, not Cost totals. Capped list rows cannot establish full population or totals; leave optional defaults to the tool.

Retrieval continuation

Recommendation list, metadata list and Service Group status return bounded pages. When the connected schema supports continuation-token, pass the returned continuationToken unchanged with the same scope, tenant, filters, ordering and page size. Keep each type’s cursor separate. Do not mix cursors between tools or scopes, parse them, or follow a URL.

Ordinary reads keep the tool defaults or explicit user limit. An explicit “more” retrieves the next page for the selected listing; “all” follows its remaining pages. Metadata lookup may follow pages to resolve an otherwise unresolved name; a partial catalog cannot establish absence or uniqueness. Stop at no token, a repeated token, no new records, an error, or an execution limit. If a token remains, or the terminal page is truncated without one, report partial coverage and the retrieved count; never claim all matches. De-duplicate stable IDs without changing returned order. Across calls the live inventory can change; this is not a point-in-time snapshot.

The 15-row inline disclosures reveal already-retrieved rows only: clicking them never calls Azure. A full table is not proof of complete retrieval, and a collapsed table is not truncation. Do not fetch more just to render supplied rows. Summary totals remain server aggregates, not sums of pages. Summary remaining-group counts describe hidden groups, not a continuation token. Prevalidation follows its discovery pages internally; its processing batches are not pages. Retrieval never selects new write targets or grants approval; use the selection safeguards.

After Advisor errors never substitute billing, Cost Management, ARM, CLI or runtime KQL. Subscription-discovery exception is defined in Scope resolution. Preserve unknown money; no reconstruction from daily rates or dates.